Step 1: Understand
Reads relevant files, searches semantically, and gathers context about the task through @-mentions and codebase indexing
Step 2: Plan
Asks clarifying questions, outlines implementation steps, and waits for approval before writing any code
Step 3: Execute
Edits files, creates new ones, runs terminal commands, and takes browser screenshots to verify visual changes
Step 4: Verify
Checks for errors, runs tests, and confirms the result. Loops back to diagnose and fix any failures automatically
0%
Developer adoption at Salesforce
0%
Engineer adoption at Brex
0%+
Fortune 500 companies using Cursor
Getting the best results from Agent mode
Start complex tasks in Plan mode so the agent outlines its approach before writing code. Be specific in your prompts: "Add Google OAuth login to the /api/auth routes using NextAuth.js" produces better results than "add login." Use @-mentions to point the agent at specific files, documentation, or URLs for reference.